ThousandML says its CQ 4-bit format preserved benchmark performance against a full-precision Qwen3.8-27B model, while its GCMP adapters matched LoRA with 55 times fewer parameters. If the results hold up beyond this test, fine-tuned models could become much cheaper to store and serve.
Watch Desk analysis
What happened
In a Hugging Face community article, ThousandML author Husam Elfadil describes two tools tested on Qwen3.8-27B: CQ, a 4-bit quantisation format, and GCMP, an adapter intended as a smaller alternative to LoRA. The post includes benchmark results and commands for reproducing them.
ThousandML reports that its 18GB CQ model showed no statistically significant difference from the 55.6GB BF16 model across nine benchmarks. It also says GCMP used 2.9 million trainable parameters and about 11MB on disk, compared with roughly 159 million parameters and 320MB for a rank-32 LoRA adapter, with no significant benchmark differences between the two.
Why it matters
The potential payoff is less storage and GPU memory for teams running many fine-tuned variants of one model. ThousandML estimates that 1,000 adapters would take about 11GB with GCMP, against 320GB with the LoRA configuration it tested. That could make serving a menu of specialised models less of a memory-juggling act.
There is a trade-off: in the reported tests, GCMP took about twice as long as LoRA per training step. The results also cover one model and one dataset, so they are a promising technical demonstration, not a universal verdict on fine-tuning.
Our read
This is worth attention because the claimed savings address a real deployment bottleneck, and the post gives readers a way to test the numbers rather than asking them to admire a headline. The practical question is whether the storage advantage survives on other models and workloads, and whether slower training is a price operators will happily pay.
What to watch
- Whether independent users reproduce the CQ and GCMP results on other models and datasets.
- Whether ThousandML publishes the pending comparisons for adapters trained on CQ and NF4 bases.
- Whether support for tensor parallelism and broader vLLM serving changes the deployment picture.
Discussion spark: Would you accept slower fine-tuning in exchange for adapters that are roughly 29 times smaller, or does training speed matter more in your workload?
Sources and evidence
- CQ and GCMP: 4-bit fine-tuning without the accuracy loss, and adapters 55× smaller than LoRA (14 November 2026, 07:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.